Jiayuan 67cd46a072 feat(agent): add context window management with token-aware compaction (#14)
* feat(agent): add context window guard to prevent token overflow

Implement token-aware context management that validates context window
size on agent initialization and provides intelligent message compaction
based on actual token usage rather than simple message count.

Key changes:
- Add context-window module with guard, token estimation, and types
- Support both "count" (legacy) and "tokens" (new default) compaction modes
- Warn when context window < 32K tokens, block when < 16K tokens
- Trigger compaction at 80% utilization, target 50% after compaction

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat(agent): add summary-based compaction using LLM

Implement intelligent compaction that uses LLM to generate summaries
of older messages instead of simply truncating them. This preserves
important context like key decisions, TODOs, and technical details.

Key changes:
- Add summarization.ts with compactMessagesWithSummary functions
- Support chunked summarization for very large histories
- Add "summary" compaction mode alongside "count" and "tokens"
- Auto-resolve API key from environment based on provider
- Graceful fallback to "tokens" mode if model/apiKey unavailable

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-30 03:46:11 +08:00

Super Multica

A multi-component architecture for distributed agent systems.

Project Structure

src/
├── agent/          # Agent module
│   ├── profile/    # Agent profile management
│   ├── session/    # Session persistence
│   └── tools/      # Agent tools (exec, process)
├── gateway/        # Gateway module
├── client/         # Client module
└── shared/         # Shared types and utilities

Getting Started

pnpm install
pnpm dev

Agent CLI

Use the agent module directly from the CLI for isolated testing.

# New sessions get a UUIDv7 ID (shown on start)
pnpm agent:cli "hello"
# [session: 019c0b0a-b111-765c-8bbd-f4149beac9c4]

# Continue a session
pnpm agent:cli --session 019c0b0a-b111-765c-8bbd-f4149beac9c4 "what did I say?"

# Or use a custom session name
pnpm agent:cli --session demo "remember my name is Alice"
pnpm agent:cli --session demo "what's my name?"

# Override provider/model
pnpm agent:cli --provider openai --model gpt-4o-mini "hi"

# Use an agent profile
pnpm agent:cli --profile my-agent "hello"

# Set thinking level
pnpm agent:cli --thinking high "solve this complex problem"

Sessions

Sessions persist conversation history to ~/.super-multica/sessions/<id>/. Each session includes:

  • session.jsonl - Message history in JSONL format
  • meta.json - Session metadata (provider, model, thinking level)

Sessions use UUIDv7 for IDs by default, providing time-ordered unique identifiers. The agent automatically handles context compaction when conversations grow too long.

Agent Profiles

Agent profiles define identity, personality, tools, and memory for an agent. Profiles are stored as markdown files in ~/.super-multica/agent-profiles/<id>/.

Profile CLI

# Create a new profile with default templates
pnpm agent:profile new my-agent

# List all profiles
pnpm agent:profile list

# Show profile contents
pnpm agent:profile show my-agent

# Open profile directory in file manager
pnpm agent:profile edit my-agent

Profile Structure

Each profile contains:

  • identity.md - Agent name and role
  • soul.md - Personality and behavioral constraints
  • tools.md - Tool usage instructions
  • memory.md - Persistent knowledge
  • bootstrap.md - Initial conversation context

Agent Tools

exec

Execute short-lived shell commands and return output.

exec({ command: "ls -la", cwd: "/path/to/dir", timeoutMs: 30000 })

process

Manage long-running background processes (servers, watchers, daemons). Output is buffered (up to 64KB) and terminated processes are automatically cleaned up after 1 hour.

# Start a background process (returns immediately with process ID)
process({ action: "start", command: "npm run dev" })

# Check process status
process({ action: "status", id: "<process-id>" })

# Read process output
process({ action: "output", id: "<process-id>" })

# Stop a process
process({ action: "stop", id: "<process-id>" })

# Clean up terminated processes
process({ action: "cleanup" })

Scripts

  • pnpm dev - Run in development mode
  • pnpm agent:cli - Run the agent CLI for module-level testing
  • pnpm agent:profile - Manage agent profiles
  • pnpm build - Build for production
  • pnpm start - Run production build
  • pnpm typecheck - Type check without emitting
Description
No description provided
Readme 270 MiB
Languages
Go 48.6%
TypeScript 42.8%
MDX 7.1%
PLpgSQL 0.4%
CSS 0.4%
Other 0.6%